Mixed finite element analysis of lognormal diffusion and multilevel Monte Carlo methods (Q507013)

Despite the background of this paper which is practical and connected to the radioactive waste disposal studies, its contents is pure mathematics. The authors study a coupled first-order system of partial differential equations with random coefficients on a bounded Lipschitz polygonal/polyhedral domain \(D\subset \mathbb R^d\), \(d=2,3\), stated in the mixed form \[ a^{-1}(\omega,\vec x)\vec q(\omega,\vec x)+ \nabla u(\omega,\vec x) = \vec g(\omega,\vec x), \quad \operatorname{div} \vec q(\omega,\vec x)=f(\omega,\vec x), \] which should be satisfied for all samples \(\omega\in\Omega\) \(P\)-almost surely. The coefficient \(a(\omega,\vec x)\) is assumed to be lognormal. In the context of steady-state flow in a porous medium, \(\vec q\) is the Darcy velocity, \(u\)~is the hydrostatic pressure and \(a\)~is the permeability. Later, an interest is on local reformulation of the initial problem, for which existence and uniqueness results are classical. For this reformulation the authors provide the reader with a finite element (FE) approximation and mixed finite element discretization. Convergence properties of these approximations are studied. Further, the mixed FE error analysis is applied to the complexity analysis of multilevel Monte Carlo methods. The crucial result is that the FE error convergence rate determines the computational cost of the ultilevel Monte Carlo. Results of numerical experiments conclude the paper.